Approval Is Not a Control

· 6 min read agentsgovernanceoperations

Navy and amber hero graphic with bold text: Approval Is Not a Control

AWS’s Well-Architected Agentic AI Lens opens its guidance on human review with a sentence worth reading twice: “Routing every agent action through human review produces rubber-stamp approvals.” That is not an outside critic. That is a hyperscaler telling architects, in its own design guidance, that the default control decays into its own absence. The same page lists insufficient reviewer context as an anti-pattern that turns review into a formality.

Most shops still treat the approval click as the control. Put a person in front of the action, log the name and the timestamp, and call the deployment governed. The log is real. The review often is not.

The numbers explain why. A LoopRails analysis of coding-agent oversight, written up in a dev.to piece on automation bias, reports that requiring plan approval cut successful attacks from about 90 percent to the 60 to 74 percent range. That sounds like a win until you read the next figure. When a bad action actually reached a human, the human stopped it only 9 to 26 percent of the time, across every oversight strategy tested. The gate helped at the level of actions. At the moment of review, the person approved anyway roughly three times out of four. The researchers call it a recognition bottleneck. The bottleneck is not attention. It is recognition: seeing the action and not recognizing what is wrong with it.

Volume makes it worse, and not in a straight line. An oversight-capacity model summarized at llm-hacking.com, based on a June 2026 arXiv paper, models the reviewer as a fatiguing resource. Realized safety comes out as an inverted U in the escalation rate. Escalate too little and danger passes unreviewed. Escalate too much and the overloaded reviewer rubber-stamps. In the simulation, escalating everything let 57 percent of danger through, while the optimal selective rate let 42 percent through. The author is explicit that this is a modeling result, not a human study. Treat it that way. The shape still matters: more approvals can mean less safety.

PAICE’s analysis of approval fatigue describes the sequence inside a real queue. Output volume rises. The queue grows faster than the reviewer pool. Time per item falls. The reviewer stops reading the reasoning and starts pattern-matching the shape of the request. Pattern-matching gets the routine ninety percent right, which trains confidence in the pattern. That confidence is what carries the reviewer past the one item that needed a real look. At every step the log shows a human approval. The audit trail is complete. The oversight is gone.

The security field is saying the same thing in harder language. ASIS Security Technology, in “The Authority Gap” (May 2026), puts it this way for aerospace and defense shops moving to an agentic workforce: when an agent processes thousands of data points per second, a human reviewer becomes a systemic vulnerability rather than a control, serving as a rubber stamp while adding latency and a false sense of security. I work in that world. Machine speed plus a human signature is not oversight. It is a liability with a name attached.

So I use a different test, and I apply it per action, not per deployment. Can this human catch this mistake in this window? If the answer is no, the approval is not a control. Stop gating the action behind a click and prevent the bad outcome by policy instead: scope the credential, cap the blast radius, block the destination, or make the action reversible. I made the placement version of this argument in The Permission Is the Policy: the grant itself has to carry the limit, because the review after it will not.

Where approval stays, tier it by reversibility and blast radius, not by action category. A reversible write inside a sandbox and an irreversible payment to an outside party are both “actions,” and routing both to the same queue is how queues die. Approve classes of routine action under policy and let them run, logged. Sample-audit the routine after the fact, where a reviewer can take real time on a few items instead of fake time on all of them. Reserve live human judgment for the irreversible and the wide-blast-radius, where a pause costs less than the error. That is also the discipline behind Stop Asking for Authority. Build the Gates.: own the gates where judgment changes the outcome, and stop pretending every decision needs a person.

The packet matters as much as the tier. A reviewer who gets the agent’s own summary of its own plan has nothing to check. The summary is written to sound right, and fluent summaries raise acceptance whether or not the action is correct. Give the reviewer an evidence packet instead: what changes, what it touches, and what undoing it costs. The actual diff, the exact recipient list, the real rows a delete will touch, the dollar amount and the payee. Evidence exposes an error. A summary hides it. I ran into the same problem from the other side in The Agent You Approved Is Not the Agent Running: approval is a snapshot, and a snapshot of a summary is two removes from the thing itself.

Then instrument the control like any other control. Track the approval rate and the median decision time per reviewer. A 98 percent approval rate at four seconds a click is not oversight. It is a log of clicks. Seed errors into the queue on a schedule and measure whether they get caught; PAICE makes the same point, that a reviewer who approves seeded errors at the same rate as clean items is performing review, not doing it. And make denial cheap. If denying an action triggers a meeting, a ticket, and a conversation about trust, nobody denies. A denial has to be blameless and fast, or the path of least resistance is always yes.

Placement still decides whether any of this survives contact with the agent itself. In The Monitor Must Live Outside the Agent’s Reach I argued that oversight an agent can edit is a suggestion. The approval path is part of that stack. If the agent can rewrite its own escalation rules, widen its own approval class, or summarize itself into a yes, the gate is decorative. A signed credential does not fix this either, which is the point I made in The Passport Is Not the Decision: a signature proves a scope was authorized. It does not prove the scope was narrow, current, or checked at the moment of action.

Approval is a control under narrow conditions: a short queue, a catchable error, real evidence, enough time, and a reviewer who can say no without paying for it. Outside those conditions, the honest move is to say so and build the control somewhere else. A gate nobody can actually work is worse than no gate, because it converts an open risk into a signed one.